Credit Card Transactions Data Adversarial Augmentation in the Frequency Domain
Mingtian Shao, Naijie Gu, Xiaoci Zhang · 2020
With the rapid development of e-commerce, a huge amount of trade data has been generated, which may be used by criminals for fraudulent transactions. In actual application, not all of the funds steal incidents can be found by the anomaly detection system in the bank because the system is not sensitive enough. A large amount of capital was taken by offenders every year, this makes a bad effect on the whole society. Thus, it becomes important to identify such fraudulent transactions in the massive commerce data. The advance of science and technology brings about the increase of data volume, as well as reveal the new application of machine learning, which provides researchers a way to detect vulnerability. Machine learning methods are widely used in the field of credit card fraud detection, but the imbalance between different categories poses an obstacle in the learning tasks. To alleviate this kind of issue, in this paper a model based on frequency domain characteristics be proposed, which is combined with a generative adversarial network (GAN) to augment minority class. In contrast to the traditional adversarial network, which only generates adversarial samples from the training data itself, this approach uses the frequency domain amplitude features of the data to generate diverse training data that matches the trend of data changes. Experimental results demonstrate classification performance is considerably improved and that over-fitting is alleviated when applying to multiple original datasets. Moreover, outperform other existing state-of the-art approaches.